search

chensnathan / YOLOF

You Only Look One-level Feature (YOLOF), CVPR2021, Detectron2
MIT License
270 stars 28 forks source link

Hi, could you provide a detailed log of the multi-scale training(R50_C5)? #27

Open x-x110 opened 3 years ago

x-x110 commented 3 years ago

Hi, could you provide a detailed log of the multi-scale training(R50_C5)?

chensnathan commented 3 years ago

Hi, sorry for the late reply.

I don't have log files for multi-scale training now. You can train it for yourself. By following the settings in the paper, you can achieve comparable results with those reported in the paper.

x-x110 commented 3 years ago

Hi, I want to run a demo using YOLOF. And, I wrote these files. However, I can't recognize any objects in the image. Could you get me some advances?

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年7月2日(星期五) 中午12:14 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

Hi, sorry for the late reply.

I don't have log files for multi-scale training now. You can train it for yourself. By following the settings in the paper, you can achieve comparable results with those reported in the paper.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

chensnathan commented 3 years ago

I don't understand that what files did you write. Could you provide more details about how you run a demo with YOLOF?

x-x110 commented 3 years ago

Hi, the logit of these files is as follows. First, following https://github.com/facebookresearch/detectron2/blob/master/GETTING_STARTED.md, we download demo.py and perdictor.py. Then, we write default params(e.g. --config-file and --input). Finally, we write the YOLOF_predictor.  Compared with default, we modifiy the checkpointer and  pre-processing of image（YOLOFCheckpointer，T.AugmentationList(build_augmentation(cfg, False))). 

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年7月11日(星期天) 晚上6:46 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

I don't understand that what files did you write. Could you provide more details about how you run a demo with YOLOF?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

chensnathan commented 3 years ago

For the demo, you can use the "DefaultPredictor" directly. Could you debug the output of the predictor? I will help with it when I got time this week.

x-x110 commented 3 years ago

Thanks!!!

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年7月12日(星期一) 晚上6:15 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

For the demo, you can use the "DefaultPredictor" directly. Could you debug the output of the predictor? I will help with it when I got time this week.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

x-x110 commented 3 years ago

Hi, we use the "DefaultPredictor" and get the result. However, I get different performances. When I modify the "SCORE_THRESH_TEST" to 0.3, we can get right, meanwhile, the Map reduces to 35.49(base 37.5). Could you give me some direction?

0.3(Map 35.49)

0.05(Map 37.5)

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年7月12日(星期一) 晚上6:15 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

For the demo, you can use the "DefaultPredictor" directly. Could you debug the output of the predictor? I will help with it when I got time this week.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

chensnathan commented 3 years ago

Hi, sorry for the late reply.

It's normal that the performance drops when you set a higher threshold (e.g., 0.3). A higher threshold means that you remove several valid predictions compared with the original setting (threshold=0.05). In YOLOF, we set the threshold as 0.05 by default.

x-x110 commented 3 years ago

Hi, maybe these pictures are not shown on GitHub. Please see the email. The problem is that I obtain many low score boxes when setting the threshold to 0.05, but I can get a clear result when setting the threshold to 0.3. But, setting the threshold to 0.3, we only get 35.49 mAP and to 0.05 get 37.5. I have put these pictures in the attachment.

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年8月3日(星期二) 晚上8:54 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

Hi, sorry for the late reply.

It's normal that the performance drops when you set a higher threshold (e.g., 0.3). A higher threshold means that you remove several valid predictions compared with the original setting (threshold=0.05). In YOLOF, we set the threshold as 0.05 by default.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

chensnathan commented 3 years ago

There exist many TPs (True Positives) between 0.05 and 0.3. Thus the mAP is lower than the original one when you set the threshold to 0.3. A detailed analysis on TPs and FPs may be helpful to understand why the performance drops.

x-x110 commented 3 years ago

Hello, when setting the threshold to 0.3, we can get the picture named 0.3.jpg(35.5mAP, true result) and to 0.05, we can get the picture named 0.05.png(37.5mAP, false result). The mAP is high but gets a false result. why? 

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年8月4日(星期三) 晚上9:06 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

There exist many TPs (True Positives) between 0.05 and 0.3. Thus the mAP is lower than the original one when you set the threshold to 0.3. A detailed analysis on TPs and FPs may be helpful to understand why the performance drops.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

chensnathan commented 3 years ago

You should check the whole validation set instead of one single image.

x-x110 commented 3 years ago

Hi, we use the coco2017 val dataset. In the attachment we submit 3 json files (improved result (our), original result (your), official result) and a simple test script. The script shows that we are able to get a good result but the detection image shows a messy frame. Why?

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年8月5日(星期四) 中午11:31 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

You should check the whole validation set instead of one single image.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

从QQ邮箱发来的超大附件

test.zip (35.69M, 无限期)进入下载页面：http://mail.qq.com/cgi-bin/ftnExs_download?k=0b6537634c35a1ca598f3dc81033561853515453075052551b545452541e505403001a5a5751561a525153515057510254500654363b64435316434d4c5a14370b&t=exs_ftn_download&code=6e7c63d7

chensnathan commented 3 years ago

I check several visualizations. There indeed exists low score bounding boxes in some images, which may be wrong predictions. While for the performance calculation, you need to do the counting for TPs, FPs, and FNs, which is more intuitive to understand why the mAP is higher with a threshold of 0.05.

BTW, you can do a visualization with different thresholds for other detectors' results. And you can get similar visualization results.

x-x110 commented 3 years ago

 Hi, I want to use your data of YOLOF(R50C5) in our paper. GFLOPs are 86 in your paper, but I got 85 when evalution the GFLOPs in Detectron2. Could you give me some adviance?

[08/23 20:06:35 d2.data.datasets.coco]: Loaded 5000 images in COCO format from /mnt/disk2/dataset/coco/annotations/instances_val2017.json [08/23 20:06:35 d2.data.build]: Distribution of instances among all 80 categories:     category    #instances       category    #instances       category    #instances   
    person      10777           bicycle    314                car      1918         
  motorcycle    367             airplane    143                bus      283         
     train      190              truck      414               boat      424         
traffic light 634          fire hydrant 101             stop sign    75           
parking meter 60                bench      411               bird      427         
      cat      202               dog      218               horse      272         
     sheep      354               cow      372             elephant    252         
     bear      71                zebra      266              giraffe    232         
   backpack    371             umbrella    407              handbag    540         
      tie      252             suitcase    299              frisbee    115         
     skis      241            snowboard    69              sports ball  260         
     kite      327          baseball bat 145          baseball gl.. 148         
  skateboard    179            surfboard    267          tennis racket 225         
    bottle      1013            wine glass  341                cup      895         
     fork      215              knife      325               spoon      253         
     bowl      623              banana    370               apple      236         
   sandwich    177              orange    285             broccoli    312         
    carrot      365             hot dog    125               pizza      284         
     donut      328               cake      310               chair      1771         
     couch      261          potted plant 342                bed      163         
dining table  695              toilet    179                tv        288         
    laptop      231              mouse      106              remote      283         
   keyboard    153            cell phone  262             microwave    55           
     oven      143             toaster    9                 sink      225         
refrigerator  126               book      1129               clock      267         
     vase      274             scissors    36              teddy bear    190         
  hair drier    11              toothbrush  57                                         
     total      36335                                                                  
[08/23 20:06:35 d2.data.dataset_mapper]: [DatasetMapper] Augmentations used in inference: [ResizeShortestEdge(short_edge_length=(800, 800), max_size=1333, sample_style='choice')] [08/23 20:06:35 d2.data.common]: Serializing 5000 elements to byte tensors and concatenating them all ... [08/23 20:06:35 d2.data.common]: Serialized dataset takes 19.15 MiB [08/23 20:06:39 fvcore.common.checkpoint]: Loading checkpoint from ../weights/YOLOF_R50_C5_1x.pth [08/23 20:06:39 fvcore.common.checkpoint]: The checkpoint state_dict contains keys that are not used by the model:   anchor_generator.cell_anchors.0 WARNING [08/23 20:06:39 fvcore.nn.jit_analysis]: Unsupported operator aten::log encountered 1 time(s) WARNING [08/23 20:06:39 fvcore.nn.jit_analysis]: The following submodules of the model were never called during the trace of the graph. They may be unused, or they were accessed by direct calls to .forward() or via other python methods. In the latter case they will have zeros for statistics, though their statistics will still contribute to their parent calling module. model.anchor_matcher [08/23 20:07:07 detectron2]: Flops table computed from only one input sample: module                              #parameters or shape    #flops     
model                                44.113M                84.517G   
  backbone                              23.455M                  66.945G   
   backbone.stem.conv1                   9.408K                   2.078G   
    backbone.stem.conv1.weight            (64, 3, 7, 7)                   
    backbone.stem.conv1.norm                                      68.352M
   backbone.res2                         0.213M                   11.75G   
    backbone.res2.0                      73.728K                  4.108G 
    backbone.res2.1                      69.632K                  3.821G 
    backbone.res2.2                      69.632K                  3.821G 
   backbone.res3                         1.212M                   16.487G 
    backbone.res3.0                      0.377M                  5.135G 
    backbone.res3.1                      0.279M                  3.784G 
    backbone.res3.2                      0.279M                  3.784G 
    backbone.res3.3                      0.279M                  3.784G 
   backbone.res4                         7.078M                   23.882G 
    backbone.res4.0                      1.507M                  5.092G 
    backbone.res4.1                      1.114M                  3.758G 
    backbone.res4.2                      1.114M                  3.758G 
    backbone.res4.3                      1.114M                  3.758G 
    backbone.res4.4                      1.114M                  3.758G 
    backbone.res4.5                      1.114M                  3.758G 
   backbone.res5                         14.942M                 12.749G 
    backbone.res5.0                      6.029M                  5.147G 
    backbone.res5.1                      4.456M                  3.801G 
    backbone.res5.2                      4.456M                  3.801G 
  encoder                              4.534M                  3.861G   
   encoder.lateral_conv                 1.049M                   0.891G   
    encoder.lateral_conv.weight          (512, 2048, 1, 1)               
    encoder.lateral_conv.bias            (512,)                         
   encoder.lateral_norm                 1.024K                   2.176M   
    encoder.lateral_norm.weight          (512,)                         
    encoder.lateral_norm.bias            (512,)                         
   encoder.fpn_conv                     2.36M                   2.005G   
    encoder.fpn_conv.weight              (512, 512, 3, 3)               
    encoder.fpn_conv.bias                (512,)                         
   encoder.fpn_norm                     1.024K                   2.176M   
    encoder.fpn_norm.weight              (512,)                         
    encoder.fpn_norm.bias                (512,)                         
   encoder.dilated_encoder_blocks       1.123M                   0.96G   
    encoder.dilated_encoder_blocks.0     0.281M                  0.24G   
    encoder.dilated_encoder_blocks.1     0.281M                  0.24G   
    encoder.dilated_encoder_blocks.2     0.281M                  0.24G   
    encoder.dilated_encoder_blocks.3     0.281M                  0.24G   
  decoder                              16.124M                  13.71G   
   decoder.cls_subnet                   4.722M                   4.015G   
    decoder.cls_subnet.0                  2.36M                    2.005G 
    decoder.cls_subnet.1                  1.024K                  2.176M 
    decoder.cls_subnet.3                  2.36M                    2.005G 
    decoder.cls_subnet.4                  1.024K                  2.176M 
   decoder.bbox_subnet                   9.443M                   8.03G   
    decoder.bbox_subnet.0                2.36M                    2.005G 
    decoder.bbox_subnet.1                1.024K                  2.176M 
    decoder.bbox_subnet.3                2.36M                    2.005G 
    decoder.bbox_subnet.4                1.024K                  2.176M 
    decoder.bbox_subnet.6                2.36M                    2.005G 
    decoder.bbox_subnet.7                1.024K                  2.176M 
    decoder.bbox_subnet.9                2.36M                    2.005G 
    decoder.bbox_subnet.10                1.024K                  2.176M 
   decoder.cls_score                     1.844M                   1.567G   
    decoder.cls_score.weight              (400, 512, 3, 3)               
    decoder.cls_score.bias                (400,)                         
   decoder.bbox_pred                     92.18K                   78.336M 
    decoder.bbox_pred.weight              (20, 512, 3, 3)                 
    decoder.bbox_pred.bias                (20,)                           
   decoder.object_pred                   23.045K                 19.584M 
    decoder.object_pred.weight            (5, 512, 3, 3)                 
    decoder.object_pred.bias              (5,)                           

[08/23 20:07:07 detectron2]: Average GFlops for each type of operators: [('conv', 86.96688461248), ('batch_norm', 0.9727329696)] [08/23 20:07:07 detectron2]: Total GFlops: 87.9±9.7

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年8月5日(星期四) 下午5:33 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

I check several visualizations. There indeed exists low score bounding boxes in some images, which may be wrong predictions. While for the performance calculation, you need to do the counting for TPs, FPs, and FNs, which is more intuitive to understand why the mAP is higher with a threshold of 0.05.

BTW, you can do a visualization with different thresholds for other detectors' results. And you can get similar visualization results.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

chensnathan commented 3 years ago

For flops calculation, we follow the steps of DETR. You can check here.

x-x110 commented 1 year ago

Sorry to bother you. The purpose of this letter is to inquire about the configuration or weight of multi-scale training (R50 or R101).  During the past two years, we was committed to solving YOLOF's NMS problem. Recently we successfully implemented the YOLOF version of NMS-Free without any additional parameters (37.1 mAP vs 37.7 mAP). But because there is no weight of multi-scale training, we can not carry out multi-scale training. After following the Settings in the paper, we can only get ~40 maps. We hope that you can provide us with a weight of multi-scale training to complete our final experiment.

------------------ 原始邮件 ------------------ 发件人: "Xx" @.>; 发送时间: 2021年8月23日(星期一) 晚上10:00 @.>;

主题: 回复： [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

Hi, I want to use your data of YOLOF(R50C5) in our paper. GFLOPs are 86 in your paper, but I got 85 when evalution the GFLOPs in Detectron2. Could you give me some adviance?

[32m[08/23 20:06:35 d2.data.datasets.coco]: [0mLoaded 5000 images in COCO format from /mnt/disk2/dataset/coco/annotations/instances_val2017.json [32m[08/23 20:06:35 d2.data.build]: [0mDistribution of instances among all 80 categories: [36m    category    #instances       category    #instances       category    #instances   
    person      10777           bicycle    314                car      1918         
  motorcycle    367             airplane    143                bus      283         
     train      190              truck      414               boat      424         
traffic light 634          fire hydrant 101             stop sign    75           
parking meter 60                bench      411               bird      427         
      cat      202               dog      218               horse      272         
     sheep      354               cow      372             elephant    252         
     bear      71                zebra      266              giraffe    232         
   backpack    371             umbrella    407              handbag    540         
      tie      252             suitcase    299              frisbee    115         
     skis      241            snowboard    69              sports ball  260         
     kite      327          baseball bat 145          baseball gl.. 148         
  skateboard    179            surfboard    267          tennis racket 225         
    bottle      1013            wine glass  341                cup      895         
     fork      215              knife      325               spoon      253         
     bowl      623              banana    370               apple      236         
   sandwich    177              orange    285             broccoli    312         
    carrot      365             hot dog    125               pizza      284         
     donut      328               cake      310               chair      1771         
     couch      261          potted plant 342                bed      163         
dining table  695              toilet    179                tv        288         
    laptop      231              mouse      106              remote      283         
   keyboard    153            cell phone  262             microwave    55           
     oven      143             toaster    9                 sink      225         
refrigerator  126               book      1129               clock      267         
     vase      274             scissors    36              teddy bear    190         
  hair drier    11              toothbrush  57                                         
     total      36335                                                                  [0m
[32m[08/23 20:06:35 d2.data.dataset_mapper]: [0m[DatasetMapper] Augmentations used in inference: [ResizeShortestEdge(short_edge_length=(800, 800), max_size=1333, sample_style='choice')] [32m[08/23 20:06:35 d2.data.common]: [0mSerializing 5000 elements to byte tensors and concatenating them all ... [32m[08/23 20:06:35 d2.data.common]: [0mSerialized dataset takes 19.15 MiB [32m[08/23 20:06:39 fvcore.common.checkpoint]: [0mLoading checkpoint from ../weights/YOLOF_R50_C5_1x.pth [32m[08/23 20:06:39 fvcore.common.checkpoint]: [0mThe checkpoint state_dict contains keys that are not used by the model:   [35manchor_generator.cell_anchors.0[0m [5m[31mWARNING[0m [32m[08/23 20:06:39 fvcore.nn.jit_analysis]: [0mUnsupported operator aten::log encountered 1 time(s) [5m[31mWARNING[0m [32m[08/23 20:06:39 fvcore.nn.jit_analysis]: [0mThe following submodules of the model were never called during the trace of the graph. They may be unused, or they were accessed by direct calls to .forward() or via other python methods. In the latter case they will have zeros for statistics, though their statistics will still contribute to their parent calling module. model.anchor_matcher [32m[08/23 20:07:07 detectron2]: [0mFlops table computed from only one input sample: module                              #parameters or shape    #flops     
model                                44.113M                84.517G   
  backbone                              23.455M                  66.945G   
   backbone.stem.conv1                   9.408K                   2.078G   
    backbone.stem.conv1.weight            (64, 3, 7, 7)                   
    backbone.stem.conv1.norm                                      68.352M
   backbone.res2                         0.213M                   11.75G   
    backbone.res2.0                      73.728K                  4.108G 
    backbone.res2.1                      69.632K                  3.821G 
    backbone.res2.2                      69.632K                  3.821G 
   backbone.res3                         1.212M                   16.487G 
    backbone.res3.0                      0.377M                  5.135G 
    backbone.res3.1                      0.279M                  3.784G 
    backbone.res3.2                      0.279M                  3.784G 
    backbone.res3.3                      0.279M                  3.784G 
   backbone.res4                         7.078M                   23.882G 
    backbone.res4.0                      1.507M                  5.092G 
    backbone.res4.1                      1.114M                  3.758G 
    backbone.res4.2                      1.114M                  3.758G 
    backbone.res4.3                      1.114M                  3.758G 
    backbone.res4.4                      1.114M                  3.758G 
    backbone.res4.5                      1.114M                  3.758G 
   backbone.res5                         14.942M                 12.749G 
    backbone.res5.0                      6.029M                  5.147G 
    backbone.res5.1                      4.456M                  3.801G 
    backbone.res5.2                      4.456M                  3.801G 
  encoder                              4.534M                  3.861G   
   encoder.lateral_conv                 1.049M                   0.891G   
    encoder.lateral_conv.weight          (512, 2048, 1, 1)               
    encoder.lateral_conv.bias            (512,)                         
   encoder.lateral_norm                 1.024K                   2.176M   
    encoder.lateral_norm.weight          (512,)                         
    encoder.lateral_norm.bias            (512,)                         
   encoder.fpn_conv                     2.36M                   2.005G   
    encoder.fpn_conv.weight              (512, 512, 3, 3)               
    encoder.fpn_conv.bias                (512,)                         
   encoder.fpn_norm                     1.024K                   2.176M   
    encoder.fpn_norm.weight              (512,)                         
    encoder.fpn_norm.bias                (512,)                         
   encoder.dilated_encoder_blocks       1.123M                   0.96G   
    encoder.dilated_encoder_blocks.0     0.281M                  0.24G   
    encoder.dilated_encoder_blocks.1     0.281M                  0.24G   
    encoder.dilated_encoder_blocks.2     0.281M                  0.24G   
    encoder.dilated_encoder_blocks.3     0.281M                  0.24G   
  decoder                              16.124M                  13.71G   
   decoder.cls_subnet                   4.722M                   4.015G   
    decoder.cls_subnet.0                  2.36M                    2.005G 
    decoder.cls_subnet.1                  1.024K                  2.176M 
    decoder.cls_subnet.3                  2.36M                    2.005G 
    decoder.cls_subnet.4                  1.024K                  2.176M 
   decoder.bbox_subnet                   9.443M                   8.03G   
    decoder.bbox_subnet.0                2.36M                    2.005G 
    decoder.bbox_subnet.1                1.024K                  2.176M 
    decoder.bbox_subnet.3                2.36M                    2.005G 
    decoder.bbox_subnet.4                1.024K                  2.176M 
    decoder.bbox_subnet.6                2.36M                    2.005G 
    decoder.bbox_subnet.7                1.024K                  2.176M 
    decoder.bbox_subnet.9                2.36M                    2.005G 
    decoder.bbox_subnet.10                1.024K                  2.176M 
   decoder.cls_score                     1.844M                   1.567G   
    decoder.cls_score.weight              (400, 512, 3, 3)               
    decoder.cls_score.bias                (400,)                         
   decoder.bbox_pred                     92.18K                   78.336M 
    decoder.bbox_pred.weight              (20, 512, 3, 3)                 
    decoder.bbox_pred.bias                (20,)                           
   decoder.object_pred                   23.045K                 19.584M 
    decoder.object_pred.weight            (5, 512, 3, 3)                 
    decoder.object_pred.bias              (5,)                           

[32m[08/23 20:07:07 detectron2]: [0mAverage GFlops for each type of operators: [('conv', 86.96688461248), ('batch_norm', 0.9727329696)] [32m[08/23 20:07:07 detectron2]: [0mTotal GFlops: 87.9±9.7

------------------ 原始邮件 ------------------ 发件人: "chensnathan/YOLOF" @.>; 发送时间: 2021年8月5日(星期四) 下午5:33 @.>; @.**@.>; 主题: Re: [chensnathan/YOLOF] Hi, could you provide a detailed log of the multi-scale training(R50_C5)? (#27)

I check several visualizations. There indeed exists low score bounding boxes in some images, which may be wrong predictions. While for the performance calculation, you need to do the counting for TPs, FPs, and FNs, which is more intuitive to understand why the mAP is higher with a threshold of 0.05.

BTW, you can do a visualization with different thresholds for other detectors' results. And you can get similar visualization results.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.